Skip to content

SYSTEM Cited by 1 source

VF Agent

VF Agent is the prototype natural-language-query layer on top of VF Match's Foundational Data Refresh (FDR) data. Built in LangGraph as a multi-agent system, VF Agent lets medical experts query the global healthcare- facility / NGO catalog using natural language — "find me orthopedic volunteer opportunities in Ghana with X-ray equipment available" — without a SQL skill prerequisite.

Stub page. Prototype, not yet production-deployed; first wiki disclosure 2026-05-20 via the Databricks + Virtue Foundation post.

Architecture

VF Agent is composed of four named sub-agents with a clear routing-supervisor shape:

                ┌──────────────────────────────────┐
        User    │ Medical Specialty Extractor      │
        query   │ (free-text → standardized        │
        ───────▶│  medical terminology)            │
                └──────────────────────────────────┘
                            │
                            ▼
                ┌──────────────────────────────────┐
                │ Multi-Agent Supervisor           │
                │ (classify intent + complexity)   │
                └──────────────────────────────────┘
                            │
                  ┌─────────┴──────────┐
                  ▼                    ▼
        ┌────────────────┐    ┌─────────────────┐
        │ Vector Search  │    │ Genie Agent     │
        │ Agent          │    │ (analytical     │
        │ (facility      │    │  queries against│
        │  discovery and │    │  structured     │
        │  search)       │    │  data)          │
        └────────────────┘    └─────────────────┘

(Source: sources/2026-05-20-databricks-virtue-foundation-medical-volunteers-72-countries)

The four sub-agents

  1. Medical Specialty Extractor — converts user free-text into standardised medical terminology (e.g. "heart surgery" → "cardiothoracic surgery"). The normalisation step gates the downstream routing decisions; fuzzy or non-standard input gets resolved to canonical taxonomy terms before the supervisor sees it.
  2. Multi-Agent Supervisor — classifies the normalised query's intent and complexity, routing to the right downstream agent. The supervisor is the canonical instance of patterns/multi-agent-supervisor-routing on the wiki.
  3. Vector Search Agent — handles facility discovery and search queries via Mosaic AI Vector Search over the embedded FDR data. "Find me hospitals in Mongolia that have orthopedic specialty" routes here.
  4. Genie Agent — handles structured-analytical queries via AI/BI Genie over the FDR Delta tables. "How many facilities in Ghana have CT scanners, broken down by region" routes here.

Why this shape

Vector search and structured-analytical query are structurally different answer-shapes:

  • Vector search: best-similarity-first retrieval — semantic proximity over an embedding space; the answer is a ranked facility list. Optimal for find-me-something-like-this queries.
  • Genie / SQL analytics: aggregate / group-by / filter against structured columns; the answer is a number, a chart, or a row set. Optimal for count / sum / breakdown queries.

A single LLM trying to do both ends up either bad at retrieval (treating everything as analytics) or bad at analytics (treating everything as similarity search). Routing per query — the supervisor's job — preserves both modes' strengths.

VF Agent is a specialisation of multi-agent architectures focused on alternative-selection routing rather than collaborative role decomposition. Compare with Claroty's three-role ER multi-agent (parse vs reason vs review collaborate on one canonicalisation task) — VF Agent's sub-agents are alternatives selected per query, not collaborators on one task.

Substrate

  • LangGraph — the agent-orchestration graph framework that hosts the supervisor + sub-agent edges.
  • Mosaic AI Vector Search — embedded-FDR-data substrate for the Vector Search Agent.
  • AI/BI Genie — embedded NL-query substrate for the Genie Agent.
  • Databricks Model Serving — substrate for the Medical Specialty Extractor + Multi-Agent Supervisor LLM endpoints.

Caveats

  • Prototype. The post explicitly frames VF Agent as "a prototype of an agent that enables experts to analyze data using natural language" — production-deployment, accuracy, latency, cost, and adoption metrics are not disclosed.
  • No supervisor-classification accuracy disclosure. The supervisor's intent / complexity classification accuracy is not reported; routing failures (e.g. a similarity-search query routed to Genie) would degrade UX silently.
  • No tool / function-calling pattern disclosed. Whether the supervisor invokes sub-agents synchronously, in parallel, or with confidence-weighted fan-out is not described.
  • Specialty-extractor terminology source (custom medical ontology vs SNOMED CT vs UMLS vs LLM-zero-shot) is not disclosed.

Seen in

Last updated · 542 distilled / 1,571 read